Pose Estimation of Randomly Organized Stator Housings using Structured Light and Harmonic Shape Contexts

This project deals with the development of a computer vision system for handling an instance of the bin picking problem. The system performs pose estimation of randomly organized stator housings with a known CAD model based on 3-D reconstruction of the bin surface. The implemented systems utilizes a structured light system based on temporally encoded binary stripes for reconstructing a tessellated point cloud describing the bin surface. The reconstructed surface is processed by diffusion based smoothing and curvature based segmentation following which surface descriptors are extracted in the form of harmonic shape context, i.e. the spherical harmonic encoding of a point-based 3-D histogram. Based on correlation between harmonic shape contexts from the model base and the scene an association graph is constructed. The search for the largest set of mutually compatible matches are approached by simulated annealing. The final pose estimation is determined by solving the orthogonal Procrustes problem for the determined point matches. Tests on the OpenGL and OpenCV based C++ implementation showed that the method was able to register simulated meshes, however, with large deteriorations given especially quantization noise. Moreover, the method was able to successfully perform pose estimation in a noise free simulated cluttered scene. The primary outcome was that the proposed method is able to perform pose estimation of randomly organized stator housing, but more consideration is needed with respect to robustly handling noise.